19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM 1015 ARTIFICIAL INTELLIGENCE: DISTINGUISHING BETWEEN TYPES & DEFINITIONS Rex Martinez* “We should make every effort to understand the new technology. We should take into account the possibility that developing technology may have im- portant societal implications that will become apparent only with time. We should not jump to the conclusion that new technology is fundamentally the same as some older thing with which we are familiar. And we should not hasti- ly dismiss the judgment of legislators, who may be in a better position than we are to assess the implications of new technology.”–Supreme Court Justice Samuel Alito 1 TABLE OF CONTENTS INTRODUCTION............................................................................................. 1016 I. WHY THIS MATTERS......................................................................... 1018 II. WHAT IS ARTIFICIAL INTELLIGENCE? ............................................... 1023 A. The Development of Artificial Intelligence ............................... 1023 B. Computer Science Approaches to Artificial Intelligence .......... 1025 C. Autonomy .................................................................................. 1026 D. Strong AI & Weak AI ................................................................ 1027 III. CURRENT STATE OF AI DEFINITIONS ................................................ 1029 A. Black’s Law Dictionary ............................................................ 1029 B. Nevada ...................................................................................... 1030 C. Louisiana................................................................................... 1032 IV. METHODS FOR WRITING A GENERAL DEFINITION ............................ 1033 A. Ambiguity & Descriptors .......................................................... 1036 B. Descriptive Definition ............................................................... 1037 * J.D. Candidate, Spring 2019, William S. Boyd School of Law, University of Nevada, Las Vegas. The author would like to thank the Nevada Law Journal staff for their edits, the facul- ty and staff at the William S. Boyd School of Law for their guidance, and Professor David McClure for his help in on this student note. 1 Brown v. Entm’t Merchs. Ass’n, 564 U.S. 786, 806 (2011) (Alito, J., concurring).
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
1015
ARTIFICIAL INTELLIGENCE:
DISTINGUISHING BETWEEN TYPES &
DEFINITIONS
Rex Martinez*
“We should make every effort to understand the new technology. We should
take into account the possibility that developing technology may have im-
portant societal implications that will become apparent only with time. We
should not jump to the conclusion that new technology is fundamentally the
same as some older thing with which we are familiar. And we should not hasti-
ly dismiss the judgment of legislators, who may be in a better position than we
are to assess the implications of new technology.”–Supreme Court Justice
I. WHY THIS MATTERS ......................................................................... 1018
II. WHAT IS ARTIFICIAL INTELLIGENCE? ............................................... 1023
A. The Development of Artificial Intelligence ............................... 1023
B. Computer Science Approaches to Artificial Intelligence .......... 1025
C. Autonomy .................................................................................. 1026
D. Strong AI & Weak AI ................................................................ 1027
III. CURRENT STATE OF AI DEFINITIONS ................................................ 1029
A. Black’s Law Dictionary ............................................................ 1029
B. Nevada ...................................................................................... 1030
C. Louisiana................................................................................... 1032
IV. METHODS FOR WRITING A GENERAL DEFINITION ............................ 1033
A. Ambiguity & Descriptors .......................................................... 1036
B. Descriptive Definition ............................................................... 1037
* J.D. Candidate, Spring 2019, William S. Boyd School of Law, University of Nevada, Las Vegas. The author would like to thank the Nevada Law Journal staff for their edits, the facul-ty and staff at the William S. Boyd School of Law for their guidance, and Professor David McClure for his help in on this student note. 1 Brown v. Entm’t Merchs. Ass’n, 564 U.S. 786, 806 (2011) (Alito, J., concurring).
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
1016 NEVADA LAW JOURNAL [Vol. 19:3
C. Prescriptive Definition .............................................................. 1038
When we say or hear the phrase “Artificial Intelligence” (or “AI”) various
things can come to mind—images from robot armies trying to extinguish hu-
manity to less intimidating images of Alexa asking someone trivia questions.
The phrase has become ubiquitous to describe countless forms of advanced
technology.2 Colloquially using a term in this way can lead to general confu-
sion, but it can be far more damaging in a legal context. Unambiguous defini-
tions are critical in the application of the law, whether in determining if a burri-
to is a sandwich,3 or if a film is considered pornography.4 With the promise of
efficiency and money, companies have begun pursuing artificial intelligence
and investing copious amounts of money into its development.5 Governments
and legislatures have certainly taken note, and some states have attempted to
outline regulations for artificial intelligence.6
There is a myriad of legal questions and problems that arise from the de-
velopment and implementation of artificial intelligence. However, without def-
initions, or at the very least a clear understanding of the concept, many re-
sources will be wasted on litigation and inept policy making.7 Certain states
have acknowledged the presence of and attempted to define AI, either in terms
of a specific function or by incorporating it into other definitions, but this only
ends up skirting the question.8 Thus, there is no general legal definition for
what constitutes AI outside of a specific application, such as in the context of
autonomous automobiles or electronic agents trading in the markets.9 However,
2 See Cristiano Castelfranchi, Alan Turing’s “Computing Machinery and Intelligence”, SPRINGER SCI. & BUS. MEDIA DORDRECHT 293, 293 (2013); James Vincent, What Counts as Artificially Intelligent? AI and Deep Learning, Explained, VERGE (Feb. 29, 2016, 3:40 PM), http://www.theverge.com/2016/2/29/11133682/deep-learning-ai-explained-machine-learning [https://perma.cc/23V8-MJQE]. 3 Associated Press, Massachusetts Judge Settles Dispute by Ruling Burrito is Not a Sand-wich, FOX NEWS (Jan. 13, 2015), http://www.foxnews.com/story/2006/11/10/massachusetts-judge-settles-dispute-by-ruling-burrito-is-not-sandwich.html [https://perma.cc/X6D4-23PE]. 4 Jacobellis v. Ohio, 378 U.S. 184, 196 (1964). 5 Matthew U. Scherer, Regulating Artificial Intelligence Systems: Risks, Challenges, Com-petencies, and Strategies, 29 HARV. J.L. & TECH. 353, 354 (2016). 6 See discussion infra Part III; see generally Assemb. Con. Res. 215, 2017-18 Leg., Reg Sess. (Cal. 2018). 7 See Scherer, supra note 5, at 359–62. 8 CAL. EDUC. CODE § 75008 (West 2018); 410 ILL. COMP. STAT. 520/6 (2018); KY. REV. STAT. ANN. § 367.680 (West 2018); OHIO REV. CODE ANN. § 4763.01 (West 2018); OHIO
REV. CODE ANN. § 4768.01 (West 2018); see discussion infra Part III. 9 See discussion infra Sections III.B–III.C.
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
Spring 2019] ARTIFICIAL INTELLIGENCE 1017
the applications of AI are expanding to a point where it will be impractical to
define it for every application.10
A general definition, therefore, can be used across fields and applications
as long as the definition is flexible and encompasses the new development of
autonomous AI. To do this, it is first necessary to untangle preconceived no-
tions of AI to determine what that term encompasses. Specifically, there needs
to be a key distinction made between complex and sophisticated programs and
systems, and systems that are capable of autonomy or human-like intelligence.
Understanding this key distinction allows for an accurate definition of AI to be
used across applications.
The utility of such definitions must be understood in order to have an ap-
plicable and workable definition.11 For instance, the definitions of single-word
nouns or short phrases can range from lengthy and complicated definitions to
one- or two-word phrases.12 With AI, it is likely to be the former. It must be
noted that even with a definition—no matter how precise or specific—that def-
inition will no doubt be the subject of litigation.13 Terms that are seemingly
well-defined and settled, such as “physical force”14 or “serious drug offense,”15
are brought to the Supreme Court for clarity on scope and application.16 I point
this out not to dissuade anyone from the importance of statutory definitions, but
to illuminate the fact that no definition is perfect. However, in this particular
case the best definition would be one flexible enough to deal with new situa-
tions, as everyday there are new uses of AI.17 Nevertheless, creating a statutory
definition can be beneficial to head off the unavoidable ambiguity courts will
face when tasked with defining AI.
The goal of this note is to first explain the invasiveness and impact that AI
has and will have in our lives, and thus, why a legal definition is necessary for
10 See discussion infra Part I. 11 Jeanne Frazier Price, Wagging, Not Barking: Statutory Definitions, 60 CLEV. ST. L. REV. 999, 1000 (2013) (“Although practitioners and academics routinely interact with statutory definitions, there has been little discussion of the functions served by those definitions or of their utility.”). 12 Id. 13 Id. at 1001 (“But the frequency of litigation around the application of legislative defini-tions belies that assumption. Within the last few years alone, the United States Supreme Court has on many occasions considered the scope and application of terms seemingly well defined by federal statute.”). 14 Johnson v. United States, 559 U.S. 133, 135 (2010). 15 McNeil v. United States, 563 U.S. 816, 817 (2011). 16 See Price, supra note 11, at 1001 (“Within the last few years alone, the United States Su-preme Court has on many occasions considered the scope and application of terms seeming-ly well defined by federal statute.”). 17 Id. at 1051 (“In a perfect world, we would expect definitions that are ‘crisp enough to ap-ply and . . . flexible enough to deal with new situations.’ ”); Scherer, supra note 5, at 373 (“Creating a working definition of AI will be difficult, to be sure, but coming up with precise legal definitions for imprecise terms is hardly a challenge unique to AI.”); see also infra Part I.
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
1018 NEVADA LAW JOURNAL [Vol. 19:3
the legal field. Further, this note will explain what AI is, and its future devel-
opments. These explanations will help us outline computer science develop-
ments and the philosophical distinction between strong AI and weak AI. The
distinction between strong AI and weak AI is useful in making distinct legal
definitions concerning AI. In addition, this note will analyze a few current ex-
amples of definitions and discern their shortcomings. Finally, this paper will
propose two possible methods for developing a definition for strong AI using
the descriptive and prescriptive methods of legal definition drafting.
I. WHY THIS MATTERS
Artificial intelligence is pervasive in our lives, whether we realize it or
not.18 One of the most public uses of AI is in driverless cars—which have been
approved for road operation in five states as well as the District of Columbia—
and there are no signs of that growing industry slowing down.19 New develop-
ments, such as these cars, have prompted an AI arms race in the private sector,
and AI could have as much as a $50 trillion economic value by 2025.20 This
race toward better and more versatile AI has called for relegations and re-
strictions.21 Wanting to regulate new technology is nothing new, but what is
surprising is the fact that so many tech industry leaders have voiced dire con-
cerns about AI.22 For example, Elon Musk stated that we are “summoning a
demon” and that AI is probably our biggest existential threat.23 People can
question whether Musk is being hyperbolic, but Stephen Hawking,24 Bill
18 See Scherer, supra note 5, at 354 (“It may not always be obvious, but we are living in the age of intelligent machines. Artificial intelligence (“AI”) permeates our lives in numerous subtle and not-so-subtle ways, performing tasks that, until quite recently, could only be per-formed by a human with specialized knowledge, expensive training, or a government-issued license.”). 19 Id.; Aaron M. Kessler, Law Left Behind as Hands-Free Cars Cruise, STARTRIBUNE (May 3, 2015, 12:21 PM), http://www.startribune.com/law-left-behind-as-hands-free-cars-cruise/3
02322781/ [https://perma.cc/3K7A-TVV7]. 20 See Scherer, supra note 5, at 354; Charles Mizrahi, The Economic Impact of AI Projected to Be Over $14 Trillion, BANYAN HILL (Jan. 24, 2019), https://banyanhill.com/economic-im
pact-ai-14-trillion/ [https://perma.cc/X6YS-BZYZ]. 21 John Frank Weaver, We Need to Pass Legislation on Artificial Intelligence Early and Of-ten, SLATE (Sept. 12, 2014, 3:53 PM), https://slate.com/technology/2014/09/we-need-to-pass
-artificial-intelligence-laws-early-and-often.html [https://perma.cc/UUF3-Q86F]. 22 Scherer, supra note 5, at 355 (“The potential for further rapid advances in AI technology has prompted expressions of alarm from many quarters, including some calls for government regulation of AI development and restrictions on AI operation.”). 23 Aileen Graef, Elon Musk: We Are ‘Summoning a Demon’ with Artificial Intelligence, UPI (Oct. 27, 2014, 7:50 AM), https://www.upi.com/Business_News/2014/10/27/Elon-Musk-We-are-summoning-a-demon-with-artificial-intelligence/4191414407652 [https://perma.cc/ U2ET-BAWX]. 24 Rory Cellan-Jones, Stephen Hawking Warns Artificial Intelligence Could End Mankind, BBC (Dec. 2, 2014), http://www.bbc.com/news/technology-30290540 [https://perma.cc/D85
7-B8LZ].
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
Spring 2019] ARTIFICIAL INTELLIGENCE 1019
Gates,25 and other scientists have expressed similar concerns.26 While some
might dismiss these worries as exaggerations, there are nevertheless issues, le-
gal or otherwise, that arise out of the use and expansion of AI.
There are also a number of legal questions that are being discussed today.27
For instance, who is liable when a self-driving car causes an accident?28 Or to
what degree can physicians let AI systems diagnose illnesses?29 Who is liable
when an AI algorithm is trying to advertise to anti-Semitic groups?30 What
happens if the AI system starts discriminating against women?31 Few states
have laws to address these issues, and if they do, the laws are limited to drones
or driverless cars.32 This lack of regulation may stem from the fact that tradi-
tional methods of regulation (tort liability, strict liability, or product licensing)
are inadequate to cover AI.33 Because AI—in at least some part—is automatic,
foreseeability and control are a major issue for liability and restrictions.34 While
these are relatively new questions, it may be difficult to foresee future uses of
AI, and thus future problems.
AI had humble beginnings. But nowadays, it has innumerable applications
that were unforeseeable when it was a nascent technology. One of the first ap-
plications of AI was its use as an opponent in a game of chess.35 Chess has
25 Peter Holley, Bill Gates on Dangers of Artificial Intelligence: ‘I Don’t Understand Why Some People Are Not Concerned’, WASH. POST (Jan. 29, 2015), https://www.washingtonpost
CMF5-2XDZ]. 26 Risks from Artificial Intelligence, CTR. FOR STUDY EXISTENTIAL RISK, https://www.cser.ac
.uk/research/risks-from-artificial-intelligence/ [https://perma.cc/CZW9-GCUR] (last visited Mar. 11, 2019). 27 Brian Hall, Top 5 Legal Issues Inherent in AI and Machine Learning, TRAVERSE LEGAL (Nov. 15, 2017), https://www.traverselegal.com/blog/top-5-legal-issues-inherent-in-ai-and-machine-learning/ [https://perma.cc/L2SD-XSM7]. 28 See Scherer, supra note 5, at 356. 29 Id. 30 Aarti Shahani, Facebook Enabled Ads Targeting Anti-Semites, NPR (Sept. 15, 2017, 5:06 AM), https://www.npr.org/2017/09/15/551163392/facebook-enabled-ads-targeting-anti-semi
tes [https://perma.cc/8VNW-WV73]. 31 Jeffrey Dastin, Amazon Scraps Secret AI Recruiting Tool that Showed Bias Against Wom-en, REUTERS (Oct. 9, 2018, 8:12 PM), https://www.reuters.com/article/us-amazon-com-jobs-automation-insight/amazon-scraps-secret-ai-recruiting-tool-that-showed-bias-against-women
-idUSKCN1MK08G [https://perma.cc/9H46-66TA]. 32 See Ryan Calo, Robotics and the Lessons of Cyberlaw, 103 CALIF. L. REV. 513, 515–16 (2015); Rachel Charney, Book Note, Can Androids Plead Automatism?, A Review of When Robots Kill: Artificial Intelligence Under the Criminal Law by Gabriel Hallevy, 73 U. TORONTO FAC. L. REV. 69, 72 (2015). 33 See Scherer, supra note 5, at 356. 34 Id. at 357. 35 Mike Murphy, An AI Computer Learned How to Beat Almost Anyone at Chess in 72 Hours, QUARTZ (Sept. 16, 2015), https://qz.com/502325/an-ai-computer-learned-how-to-beat
clearly defined rules and a finite number of moves, so a computer can run
through all of the possibilities and pick the best one.36 But we have come a long
way since Deep Blue beat Kasparov in chess over two decades ago.37 For ex-
ample, I.B.M.’s Watson destroyed Jeopardy champions,38 and more recently,
Google’s DeepMind beat the world champion in Go, which is considered to be
the most complex game.39 However, the capabilities of artificial systems are
quickly going beyond games.
AI is no longer limited to computational or statistical tasks and opera-
tions.40 AI is now being used to make investments on Wall Street, and it is ca-
pable of making 193,000 trades a day.41 Furthermore, the consumer behavior
predictions used by online retail marketers are generated by an AI algorithm
with great precision.42 AI is used in the medical field to predict heart attacks
and heart disease or to detect Alzheimer’s, and more recently, AI is being used
to spot colorectal cancer tumors before they become malignant with 86 percent
accuracy.43 AI is also extremely good at computer tasks and can not only en-
hance pictures by increasing a photo’s resolution,44 but it can create realistic
fake people to fill those photos.45 Even tasks that we assume require human
36 Garry Kasparov, Garry Kasparov: There’s No Shame in Losing to a Machine, FORTUNE (Sept. 25, 2017), http://fortune.com/2017/09/25/garry-kasparov-chess-strategy-artificial-intel
ligence-ai/ [https://perma.cc/JV7W-EDT6]. 37 Id. 38 John Markoff, Computer Wins on ‘Jeopardy!’: Trivial, It’s Not, N.Y. TIMES (Feb. 16, 2011), http://www.nytimes.com/2011/02/17/science/17jeopardy-watson.html?pagewanted=a
ll [https://perma.cc/AZ7R-2EC4]. 39 Nature Video, The Computer That Mastered Go, YOUTUBE (Jan. 27, 2016), https://www.
youtube.com/watch?v=g-dKXOlsf98 [https://perma.cc/2MZ4-ZXFY]; see also Vincent, su-pra note 2 (describing additional AI abilities now being employed by big tech companies). 40 See Vincent, supra note 2. 41 Joe Ciolli, The Stock Market’s Robot Revolution is Here, BUS. INSIDER AUSTL. (Oct. 21, 2017, 9:02 AM), http://www.businessinsider.com.au/ai-powered-equity-etf-stock-aieq-marke
t-robot-revolution-2017-10 [https://perma.cc/9TLX-45C2]. 42 Meet Genie, GREY JEAN TECHS., http://gjny.com/meet-genie/ [https://perma.cc/D5GR-FE
PP] (last visited Mar. 12, 2019). 43 Drew Harwell & Carolyn Y. Johnson, In Our Eyes, Google’s Software Sees Heart Attack Risk, WASH. POST (Feb. 19, 2018), https://www.washingtonpost.com/news/the-switch/wp/20
redirect=on&utm_term=.3dea14a86c12 [https://perma.cc/W9W3-RMCA]; Swapna Krishna, An AI Detected Colorectal Cancer with 86 Percent Accuracy, ENGADGET (Oct. 30, 2017), htt
ps://www.engadget.com/2017/10/30/ai-colorectal-cancer-detection/ [https://perma.cc/W23S-DHCP]; Press Release, Radiological Soc’y of N. Am., Artificial Intelligence May Aid in Alzheimer’s Diagnosis (July 6, 2016), https://press.rsna.org/timssnet/media/pressreleases/14
_pr_target.cfm?ID=1890 [https://perma.cc/QKB2-F6FB]. 44 Camille Charluet, This New AI Can Make Your Low Resolution Photos Great Again, THE
NEXT WEB (Oct. 31, 2017), https://thenextweb.com/artificial-intelligence/2017/10/31/ai-can-make-low-resolution-photos-great-again/ [https://perma.cc/RN9H-R7QX]. 45 Roger Cheng, AI Can Create Fake Celebrity Photos so Real It’s Scary, CNET (Oct. 30, 2017, 9:05 AM), https://www.cnet.com/news/ai-can-create-fake-celebrity-photos-so-real-its-scary/?part=propeller&%3Bsubj=news&%3Btag=link [https://perma.cc/3SEW-BSDY].
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
Spring 2019] ARTIFICIAL INTELLIGENCE 1021
creativity are being tackled by AI—such as writing horror stories,46 creating
music videos,47 or developing black metal albums.48 While these applications
might seem presumably helpful and not harmful, the applications do stop there.
There are also a fair share of science fiction applications already in the
works, including: mind reading,49 predicting one’s death,50 creating computer
voice generations that are indistinguishable from humans,51 and conducting
surveillance for the Pentagon via drones and satellites.52 Last but not least, AI
systems are being programed to be able to build other AI machines.53 The uses
and functions of AI by no means end there, but it is becoming increasingly
clear that AI systems will be able to do most tasks that humans can do, and do
them better. It is also difficult to predict future uses for AI, especially since AI
systems will most likely be the ones inventing new uses.54
46 Swapna Krishna, AI Can Write Surprisingly Scary and Creative Horror Stories, ENGADGET (Oct. 31, 2017), https://www.engadget.com/2017/10/31/shelley-ai-writes-horror-stories-on-twitter/ [https://perma.cc/8YAT-3E53]. 47 Jon Fingas, Intel AI Helped Create a Music Video, ENGADGET (Oct. 28, 2017), https://ww
NT]. 48 Mark Austin, A.I. Bots Just Dropped a Black Metal Album that Will Make Your Head Ex-plode, DIGITAL TRENDS (Dec. 5, 2017, 10:04 AM), https://www.digitaltrends.com/music/dad
abots-ai-death-metal-album/ [https://perma.cc/2CG6-ESGK]. 49 Catherine Clifford, Japanese Scientists Just Used A.I. to Read Minds and It’s Amazing, CNBC (Jan. 8, 2018, 11:21 AM), https://www.cnbc.com/2018/01/08/japanese-scientists-use-artificial-intelligence-to-decode-thoughts.html [https://perma.cc/QKX2-H4AR]. 50 Catherine Clifford, These Scientists Are Using A.I. to Predict When People Will Die for Better Access to Medical Care, CNBC (Jan. 25, 2018, 9:15 AM), https://www.cnbc.com/201
.cc/DUY4-VWMQ]. 51 Dave Gershgorn, Google’s Voice-Generating AI is Now Indistinguishable from Humans, QUARTZ (Dec. 26, 2017), https://qz.com/1165775/googles-voice-generating-ai-is-now-indisti
nguishable-from-humans/?utm_source [https://perma.cc/HGY2-GFHN]. 52 Gregory C. Allen, The Pentagon is Using AI to Fight ISIS but It’s Not Quite the ‘Termi-nator’, CNN (Dec. 30, 2017, 10:01 AM), https://www.cnn.com/2017/12/29/opinions/pentag
ma.cc/AN48-BBQB]. 54 Id. Just the period between the first draft of this paper in March 2018 and the second draft in July 2018, there have been impressive developments. AI can now perform a variety of human tasks. See, e.g., Aaron Brown, The Creepy AI that Can Predict the Future: Machine That Anticipates Your Movements Several Minutes in Advance Could Pave the Way for Next-Level Big Brother Surveillance, DAILYMAIL.COM (June 15, 2018, 8:58 AM), http://www.dail
ymail.co.uk/sciencetech/article-5847767/Creepy-AI-predict-moves-advance-lead-level-Big-Brother-surveillance.html [https://perma.cc/FX8J-8ADE] (noting AI can predict human’s movements before they do them); Andrew Liszewski, MIT’s New AI-Powered Software Can Extract Individual Instruments from Videos with a Single Click, GIZMODO (July 5, 2018, 4:40 PM), https://gizmodo.com/mits-new-ai-powered-software-can-extract-individual-ins-182737
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
1022 NEVADA LAW JOURNAL [Vol. 19:3
Given the plethora of applications of AI, it seems like we should have a
clear definition of what qualities make all of these “AI.” What was once con-
sidered AI no longer qualifies as artificially intelligent technology and applica-
tions advance. However, despite this evolution, scientists, programmers, phi-
losophers, and lawyers do not have a clear understanding of what AI is.
Attempting to define AI is nothing new, but what is different now is the neces-
sity of a definition, and specifically a legal definition. In order to be able to de-
fine it, however, the legal field needs to understand what AI is, as well as what
makes it unique from other technological advances, including what is currently
being called “AI.”
2032 [https://perma.cc/J35Q-69DE] (noting AI can extract a single instrument from a song); Sam Machkovech, This Wild AI-Generated Film Is the Next Step in “Whole Movie Puppet-ry,” ARS TECHNICA (June 11, 2018, 9:01 AM), https://arstechnica.com/gaming/2018/06/this-wild-ai-generated-film-is-the-next-step-in-whole-movie-puppetry/ [https://perma.cc/34LP-J5
VT] (noting AI can create films); Andy Meek, It’s a Little Scary How Smart Google’s DeepMind Just Got, BGR (June 15, 2018, 2:58 PM), https://bgr.com/2018/06/15/google-dee
pmind-render-3d-objects-from-2d-images/ [https://perma.cc/DR78-D4ZT] (noting AI can crush players in video games); Stephanie Mlot, Artificial Intelligence Can Sense Humans Through Walls, GEEK.COM (June 14, 2018), https://www.geek.com/tech/artificial-intelligenc
e-can-sense-humans-through-walls-1743205/ [https://perma.cc/SEE2-4PRF] (noting AI can sense humans through walls); Jack Morse, Watch Out Google Duplex: Microsoft Just De-moed Its Own AI Having a Full-On Phone Call, MASHABLE (May 22, 2018), https://mashabl
H] (noting AI systems can now make full on phone calls); Angela Moscaritolo, IBM Artifi-cial Intelligence Takes on Human Debate Champs, PCMAG.COM (June 19, 2018, 12:25 PM), https://www.pcmag.com/news/361938/ibm-artificial-intelligence-takes-on-human-debate-champs [https://perma.cc/MPE3-KBYJ] (noting AI can debate people); Brian Resnick, How Data Scientists Are Using AI for Suicide Prevention, VOX (June 9, 2018, 7:22 AM), https://w
isis-text-line-data-science [https://perma.cc/85BX-SMNY] (noting AI can help prevent sui-cide); Phil Stewart, Deep in the Pentagon, a Secret AI Program to Find Nuclear Missiles, REUTERS (June 5, 2018, 3:07 AM), https://www.reuters.com/article/us-usa-pentagon-missiles
CN1J114J [https://perma.cc/3HCM-KQEW] (noting AI can find hidden nuclear missiles); James Vincent, This Japanese AI Security Camera Shows the Future of Surveillance Will Be Automated, VERGE (June 26, 2018, 7:31 AM), https://www.theverge.com/2018/6/26/174790
TGY-88C9] (noting AI can provide video surveillance); Kyle Wiggers, Google’s DeepMind Developed an IQ Test for AI Models, VENTUREBEAT (July 11, 2018, 10:52 AM), https://vent
ma.cc/A9FJ-D2DW] (noting AI can even develop a test to measure other AIs); Kyle Wig-gers, Google’s DeepMind Develops AI That Can Render 3D Objects from 2D Pictures, VENTUREBEAT (June 14, 2018, 12:13 PM), https://venturebeat.com/2018/06/14/googles-deepmind-develops-ai-that-can-render-3d-objects-from-2d-pictures/ [https://perma.cc/6BXQ
-4V4N] (noting AI can render 3D objects from 2D pictures). There will be countless other applications by the time this paper even goes to print.
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
Spring 2019] ARTIFICIAL INTELLIGENCE 1023
II. WHAT IS ARTIFICIAL INTELLIGENCE?
“What is Artificial Intelligence?” is not only a difficult question in and of
itself, but it is particularly difficult because it is not clear who can or should an-
swer the question.55 Philosophers, computer scientists, and cognitive scientists
are just a few of the groups who have attempted to answer this very question.56
Although they disagree about what constitutes AI, they tend to acknowledge
the relevance and significance of the question.57 While this question may not be
any easier to answer, it does raise the consideration of whether the legal field
even cares what intelligence is in this context. Although, the court does care
about intelligence when pertaining to people.58 A legal definition may not turn
on whether something qualifies as intelligent, but it is certainly a starting point.
A. The Development of Artificial Intelligence
Before there can be suitable regulation, legislation, or implementation,
there must be a definition-based conceptual understanding.59 The hang-up in
defining AI is, in large part, because of the ambiguity of “intelligence.”60 We
recognize intelligence in ourselves, so we link our understanding of intelligence
to human characteristics.61 One understanding of intelligence is tied to the abil-
ity to perform intellectual tasks.62 However, as technology advances, so too do
the tasks computers can accomplish.63 As machines accomplish more tasks, we
tend not to consider them as reaching intelligence, but instead we move the
threshold of intelligence farther away and then treat that specific task as unin-
dicative of intelligence.64 The history of artificial machine development has
certainly made this point clear.
Intelligent machines have been a serious focus of research since the mid-
1950s, when computer scientists began developing programs that could solve
55 See Scherer, supra note 5, at 373, 396. 56 See generally STUART J. RUSSELL & PETER NORVIG, ARTIFICIAL INTELLIGENCE: A
MODERN APPROACH xii (3d ed. 2016); John R. Searle, Minds, Brains, and Programs, BEHAV. & BRAIN SCIS. 417, 417 (1980); John E. Laird et al., A Standard Model of the Mind: Toward a Common Computational Framework Across Artificial Intelligence, Cognitive Science, Neuroscience, and Robotics, AI MAGAZINE, Winter 2017, at 13, 14. 57 See generally sources cited supra note 56. 58 Brady v. United States, 397 U.S. 742, 756 (1970); Boykin v. Alabama, 395 U.S. 238, 242 (1969) (holding that for guilty pleas to count the court must find that the defendant made it intelligently). This is one of the many examples where intelligence is a critical condition. 59 See generally Scherer, supra note 5, at 359; John McCarthy, What is Artificial Intelli-gence? STAN. FORMAL REASONING GRP. (Nov. 12, 2007), http://www-formal.stanford.edu/jm
c/whatisai/node1.html [https://perma.cc/PR7N-YGD4]. 60 See Scherer, supra note 5, at 359 (“The difficulty in defining artificial intelligence lies not in the concept of artificiality but rather in the conceptual ambiguity of intelligence.”). 61 Id. at 359–60; see generally McCarthy, supra note 59. 62 Scherer, supra note 5, at 360. 63 Id. at 360–61. 64 Id. at 361.
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
1024 NEVADA LAW JOURNAL [Vol. 19:3
problems previously thought to require intelligence.65 From there, computer
scientists developed systems with the ability to play chess, action-plan, sched-
ule tasks, and other complex tasks.66 However, it became clear that these sys-
tems were developed to mimic the behavior of human intelligence, not neces-
sarily to exercise human intelligence.67 And this clarity highlights the
distinction between sophisticated programs and machines actually capable of
thinking and decision making.68 Despite this distinction, researchers continued
to develop more sophisticated machines capable of data-mining or creating
train and airline schedules.69 As machines became more capable, the question
still remained whether this was human intelligence.70
The Turing test has been a common method used to ascertain if human in-
telligence has been reached.71 In the test, a person has a conversation with an
unknown thing on the other end, either an AI system or an actual person.72 If
the person is conversing with a machine but thinks he or she is communicating
with a person, then the machine is said to exhibit at least some intelligence.73
This test is by no means conclusive, but it is still a common test to run. Alt-
hough recently, it is worth mentioning that Google’s AI has more or less con-
quered the Turing test.74 Still, if the question is whether or not something has
intelligence, there needs to be some form of measurement.75
The Turing test has withstood the advancement of technology because it
forces the computer to imitate human behavior, giving it a strong anthropocen-
tric bias.76 For instance, what we consider simple questions—like “what is your
oldest memory?” or “what was your most painful moment?”—quickly expose
that one is speaking to a machine.77 There are other assessments and thought
experiments that highlight the complication of discerning a machine from a
65 ARLINDO OLIVEIRA, THE DIGITAL MIND: HOW SCIENCE IS REDEFINING HUMANITY 87–88 (2017). 66 Id. at 88. 67 Id. 68 See discussion infra Section II.D; see generally Searle, supra note 56, at 417. 69 See OLIVEIRA, supra note 65, at 89. 70 Id. 71 Id. at 90; see also RUSSELL & NORVIG, supra note 56, at 2. 72 OLIVEIRA, supra note 65, at 90. 73 Id. 74 David Gewirtz, Google Duplex Beat the Turing Test: Are We Doomed?, ZDNET (May 14, 2018, 6:29 PM), https://www.zdnet.com/article/google-duplex-beat-the-turing-test-are-we-doomed/ [https://perma.cc/6EL8-HYQS]; Andrew Tarantola, Pretty Sure Google’s New Talking AI Just Beat the Turing Test, ENGADGET (May 8, 2018), https://www.engadget.com/
EC-53DW]. 75 See generally HANDBOOK OF INTELLIGENCE 3 (Robert J. Sternberg ed., 2000). 76 See OLIVEIRA, supra note 65, at 90–91. 77 Id. at 91.
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
Spring 2019] ARTIFICIAL INTELLIGENCE 1025
human.78 Nevertheless, the central focus of AI research has been in determining
and mimicking human intelligence.79 There are countless advancements and
techniques that aim to reach the level of human intelligence,80 such as through
the use of pattern recognition, complex communication, and machine learn-
ing.81 However, all of this research and development has led to varying defini-
tions of what qualifies as AI, each of which focus on different attributes that are
deemed to be “intelligence.”
B. Computer Science Approaches to Artificial Intelligence
Stuart Russell and Peter Norvig outline various definitions of “artificial in-
telligence” that fall into four categories: acting humanely, thinking humanely,
thinking rationally, and acting rationally.82 The goal of these definitions is to
point to the attribute that necessarily signifies intelligence.83 First, the Turing
test is meant to demonstrate that if an AI system is acting human, then it is in-
telligent. This is an operational definition of “intelligence.”84 While acting hu-
manly may be a sign of intelligence, it is certainly not the only condition.85 The
cognitive modeling approach focuses on whether the machine thinks humanly.
Under this approach, intelligence is a product of how humans think, and thus,
an artificial intelligence system must match the human-mode of thinking in or-
der to be considered intelligent.86 Comparatively, thinking rationally involves
following the logician tradition of formalizing informal knowledge into logical
notation, which is rule following.87 Finally, the rational agent approach focuses
on acting rationally. Under that approach, an agent is one who perceives and
acts, and to act rationally is to act “so as to achieve one’s goals, given one’s be-
liefs.”88
These approaches are by no means exclusive of one another, and each ap-
proach has its faults and counter arguments as to why the other is not intelli-
78 See generally RUSSELL & NORVIG, supra note 56, at 1–4; JOHN SEARLE, MINDS, BRAINS
AND SCIENCE 28–41 (1984); Searle, supra note 56, at 417. 79 See generally OLIVEIRA, supra note 65, at 87–97. 80 Id. at 94–95. 81 See ERIK BRYNJOLFSSON & ANDREW MCAFEE, THE SECOND MACHINE AGE 17–18, 20, 22 (2014). 82 See RUSSELL & NORVIG, supra note 56, at 2–4. 83 See id. 84 Id. at 2. 85 See id. at 3. 86 Id. However, there is still the issue of figuring out how one thinks. Castelfranchi, supra note 2, at 298 (noting that in an interview Turing said, “The whole thinking process is still rather mysterious to us, but I believe the attempt to make a thinking machine will help us greatly in finding out how we think ourselves.”). 87 See RUSSELL & NORVIG, supra note 56, at 4. 88 See STUART J. RUSSELL & PETER NORVIG, ARTIFICIAL INTELLIGENCE: A MODERN
APPROACH 7 (1st ed. 1995).
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
1026 NEVADA LAW JOURNAL [Vol. 19:3
gence.89 Acting humanly can be accomplished by mimicking human respons-
es.90 Figuring out how we think is difficult, but it seems as if three separate
methods of thinking could still reach the level of intelligence.91 Humans often-
times do not think rationally.92 And acting rationally is circular because to act
toward goals, one must have beliefs, and to have beliefs one must be intelligent.
However, modern AI is based on the rational agent’s theory outlined by John
von Neumann.93 Although there is no settled characteristic or mode of intelli-
gence, these definitions highlight what we consider attributes of intelligence,
and it is these attributes that legal definition should incorporate.94 Each of these
definitions, whether explicitly stated or not, point to a critical attribute of intel-
ligence: autonomy. Each of these approaches acknowledges autonomy, regard-
less of which is indicative of intelligence.
C. Autonomy
The most distinct characteristic of intelligence is autonomy.95 Autonomy is
considered to be a condition of thinking.96 This autonomy allows for the attrib-
ution of thinking or intelligence to children and animals, but not machines, as
machines just do as they are told.97 However, as machines are becoming more
advanced, AI systems are beginning to exhibit and exercise autonomy—making
decisions free from outside input.98 This is by design, as machine learning sys-
tems need to be free from human input, otherwise AI systems simply could not
learn or function as intended.99 To be clear, there was human input in the form
of the initial programming, but the purpose of the system is to learn and to pro-
vide unprogrammed outputs.
There are multiple benefits to AI machines having autonomy. Humans tend
to think in a certain way, a way in which machines may not be bound by. Thus,
AI systems are able to come up with unconventional solutions that humans
89 See RUSSELL & NORVIG, supra note 56, at 1–2. 90 See id. at 2. 91 Id. at 3 (describing introspection, psychological experiments, and brain imaging as three separate methods of thinking). 92 See id. at 2 n.1. 93 Steve Omohundro, A Turning Point in Artificial Intelligence, in WHAT TO THINK ABOUT
MACHINES THAT THINK 12 (John Brockman ed., 2015). 94 See Scherer, supra note 5, at 360. The legal field need not, and probably should not, be the one that determines what approach is the most indicative of intelligence. What matters is that when creating a statutory definition, for any word or phrase, that the definition is as ac-curate and functional as it needs it be. 95 Id. at 363. 96 See Castelfranchi, supra note 2, at 298. 97 Id. 98 Torben Friehe, How Much Autonomy Is Too Much for AI?, NEXT WEB (Nov. 18, 2017), https://thenextweb.com/contributors/2017/11/18/much-autonomy-much-ai/ [https://perma.cc/
TEW3-JSBK]. 99 Id.; Murphy, supra note 35; Vincent, supra note 2.
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
Spring 2019] ARTIFICIAL INTELLIGENCE 1027
cannot come up with.100 Currently, AI machines acting unpredictably means
that they design something in a unique way or that they perform an unexpected
move in chess or Go.101 In the future, however, AI systems’ unpredictability
will be able to come up with solutions not only to problems we cannot solve,
but also to problems we are not even aware of, and this will happen more fre-
quently.102 This is one of the major appeals of these types of systems: the abil-
ity to learn and surpass our problem-solving abilities.103 Therefore, we need AI
systems to have some autonomy, which makes the distinction between autono-
mous decision making systems and non-autonomous systems even more criti-
cal.
Yet, with AI systems acting autonomously, there is going to be a question
about control.104 Building autonomous AI is the goal, which means that devel-
opers could lose control of the systems they create.105 Of course developers are
the ones who program the objectives of the system; however, objectives could
be vague or ambiguous and the AI system could use undesirable means to carry
out that objective.106 Moreover, AI systems might not want to give up con-
trol.107 These are a few potential considerations that highlight the distinction
between systems capable of making their own decisions and other systems that
function in prescribed parameters. Because of the potential problems, it is im-
portant to distinguish between autonomous and non-autonomous AI.
D. Strong AI & Weak AI
There is a helpful distinction from the field of philosophy of mind that
sheds light on this difficult concept: strong AI and weak AI. Although, it
should be mentioned that like most, if not all, philosophical theories, there are
people who disagree. However, this distinction is useful for semantic and
pragmatic purposes. Weak AI is the concept that whatever the program is
meant to do, it is merely trying to replicate or duplicate that function, and for
most tasks that is sufficient.108 Whereas strong AI is an actual instantiation of
that thing, which in this case is intelligence.109 Simply put, weak AI simulates,
whereas strong AI just is.
100 See Scherer, supra note 5, at 365. 101 Id.; Nature Video, supra note 39. 102 See Scherer, supra note 5, at 365 (“The AI’s solution thus may not have been foreseeable to a human—even the human that designed the AI . . . . AI behavior will crop up with in-creasing frequency and that the unexpectedness of AI behavior will rise significantly.”). 103 Id. 104 Id. at 366; see also Friehe, supra note 98. 105 See Scherer, supra note 5, at 367. 106 Id. 107 Id. at 368. 108 Searle, supra note 56, at 417. 109 Id.
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
1028 NEVADA LAW JOURNAL [Vol. 19:3
Weak AI is nothing new, it is the simulation of decision making. For ex-
ample, the system that beat Kasparov at chess was calculating the best possible
outcomes and making moves according to an algorithm.110 Here, we can say
that the system was imitating intelligence by “thinking” about which move to
make and then making it.111 The system is actually playing chess, but it is not
actually thinking or deliberating about which move to make like people
might.112 This is far different than strong AI, which is closer to decision making
based on intelligence. Google’s Deepmind system utilizes machine learning to
learn from experience and makes autonomous decisions much like what we
consider intelligence.113 Thus, it is strong AI, or at least close.114
There are a lot of questions about how fine the line is between strong AI
and weak AI, as well as questions about what qualifies as either form and how
we would know. But this distinction still highlights an important and useful dif-
ference. Colloquially, our typical use of the phrase “artificial intelligence” does
not differentiate between strong AI and weak AI, and the two are conflated and
treated equally,115 which could mean equal legal treatment. This distinction
matters because there is a difference in autonomy, which could result in une-
qual legal treatment. An analogous legal example would be the difference be-
tween treatment of a minor and of an adult. Parents can be liable for the actions
of their minor children, whereas parents are not liable for the actions of their
adult children.116 Or, in the case of criminal justice, minors are usually given
lesser sentences than adults.117 Similarly, weak AI systems (i.e., systems where
the developer has control) would make the developer liable just as a parent
would be liable for the actions of a minor. Whereas strong AI may be treated as
an adult, weak AI would be treated as a minor. To be clear, strict liability may
be the best solution for both strong AI and weak AI, but the distinction is still
helpful. We would still need the distinction between adults and minors even if
parents were strictly liable for their children. Therefore, my argument is that we
need a legal definition that both separates the weak AI and strong AI, regard-
less of whether the legal system treats them similarly or differently.
110 Kasparov, supra note 36. 111 See Murphy, supra note 35. 112 This would be the “thinking humanly.” See RUSSELL & NORVIG, supra note 56, at 3. 113 James Vincent, DeepMind’s Go-Playing AI Doesn’t Need Human Help to Beat Us Any-more, VERGE (Oct. 18, 2017, 1:00 PM), https://www.theverge.com/2017/10/18/16495548/de
epmind-ai-go-alphago-zero-self-taught [https://perma.cc/RD2Y-Q3RF]. 114 It may be difficult to tell when the hard AI threshold met, yet that does not undermine that the distinction is still there. If there is a legal difference, say in liability, between strong AI and weak AI, this threshold question will become significant as attorneys will argue for the one that best suits them. 115 See Vincent, supra note 2; see also discussion supra Part I. 116 See Larry Cunningham, A Question of Capacity: Towards a Comprehensive and Con-sistent Vision of Children and Their Status under Law, 10 U.C. DAVIS J. JUV. L. & POL’Y 275, 321, 333 (2006); Arthur T. Spence, Parental Liability, 309 INS. L.J. 787, 788 (1948). 117 Cunningham, supra note 116, at 277, 279.
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
Spring 2019] ARTIFICIAL INTELLIGENCE 1029
This section covered several fundamental conceptions and attributes of AI,
which in our current usage may or may not be part of what people mean when
they say “artificial intelligence.” The strong AI/weak AI distinction is useful in
teasing out one conception of AI from another: complex algorithmic systems
and systems that exhibit autonomous intelligence and decision making. Defin-
ing strong AI will necessarily define weak AI because setting parameters for
the former will exclude the latter. Many jurisdictions already incorporate weak
AI in their statutes because it often falls under computer codes, programs, soft-
ware, or algorithms, and thus, is already covered by some definition or provi-
sion if necessary.118 However, because of the fundamental difference between
strong AI and weak AI, a distinct legal definition is needed for strong AI. Thus,
the distinction and attributes mentioned above should be considered when for-
mulating a legal definition.
III. CURRENT STATE OF AI DEFINITIONS
Some jurisdictions that have been proactive and have defined AI in one
form other another.119 However, these definitions either lack a fundamental un-
derstanding or are entirely too broad as to provide any help at all. Furthermore,
if a jurisdiction even has a definition it will likely limit the definition to specific
contexts.120 There are many different places where a lawyer or judge can look
to for guidance when attempting to discern if a system qualifies as AI. A dic-
tionary is a common source,121 as well as statutes from other jurisdictions;122
so, it can be useful to look at other sources to determine how other jurisdictions
are currently defining artificial intelligence.
A. Black’s Law Dictionary
Black’s Law Dictionary is a commonly referred to source of terms and def-
initions. The definition of artificial intelligence in Black’s Law Dictionary is
“software used to make computers and robots work better than humans. The
systems are rule based or neutral networks. It is used to help make new prod-
ucts, robotics, human language understanding, and computer vision.”123 Signif-
icant problems exist with this definition, which run contrary to understanding
what AI is considered today, especially when juxtaposed with machine learn-
118 See sources cited supra note 8. 119 See supra Part III. 120 See supra Part III; see also NEV. REV. STAT. § 482A.020 (repealed 2013) (limiting the definition to autonomous vehicles). 121 See Artificial Intelligence, MERRIAM-WEBSTER, https://www.merriam-webster.com/dicti
onary/artificial%20intelligence?utm_campaign=sd&utm_medium=serp&utm_source=jsonld [https://perma.cc/ZM9X-EAKG] (last visited Mar. 15, 2019) (defining artificial intelligence as “the capability of a machine to imitate intelligent human behavior”). 122 See infra Sections III.B–III.C. 123 Artificial Intelligence, THE LAW DICTIONARY, https://thelawdictionary.org/artificial-intell
igence/ [https://perma.cc/5BPN-YBTE] (last visited Mar. 15, 2019).
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
1030 NEVADA LAW JOURNAL [Vol. 19:3
ing.124 However, this definition limits AI to only two forms: rule based or neu-
ral networks.125 So, for something to qualify as AI under the first two prongs,
it: (1) must be software, which does not seem controversial, although what
qualifies as software might carry with it a loaded meaning or ambiguity;126 and
(2) the systems must be rule based or neural networks.127 While this is a good
start that provides some guidance, the latter part of the definition is unhelpful.
Black’s characterizes AI as a tool, which in and of itself is not wrong. The
primary focus of development is to better our lives or to become more efficient
in some way. However, when contrasted to what we currently call “tools,” such
as hammers, books, or Westlaw, these tools seem incomparable to a software
that acts on its own. It would be wrong to say that there are not autonomous
tools in some form (e.g., automated factories), but there is a difference between
following pre-registered actions to construct a car and learning to make better
autonomous decisions on what stock to invest in.128 Put another way, it does
not seem like a tool if it is doing the entire job on its own. In addition, the latter
part of the definition is just too broad. It includes digital watches because those
are better than people at tracking time and are rule-based software. But digital
watches are far from AI. It is the last part of the definition which confuses and
misleads.
The phrase “used to help make new products, robotics, human language
understanding, and computer vision,” is also problematic as it makes it seem as
if the point of AI is limited to these things. This definition binds the qualifica-
tion to just these few uses. This is problematic because AI’s uses are by no
means limited, and thus, the definition focuses on narrow uses. Primarily, this
definition ignores the autonomy. These are just a few of the many problems
that arise out of this definition. While Black’s Law Dictionary is neither a stat-
ute nor binding precedent, it is a common tool used by the legal community and
could be a place that lawyers, judges, or policy makers may look to for guid-
ance; however, in this case, that guidance is unsatisfactory.
B. Nevada
Nevada does not have a standalone definition for AI, but they did attempt
to define it in terms of a specific application: autonomous vehicles.129 While
this statute was ultimately repealed, it demonstrates both an approach to defin-
ing AI, and a significant modern development in AI that states will have to con-
124 See discussion supra Part I. 125 See Artificial Intelligence, supra note 123. 126 Software, BLACK’S LAW DICTIONARY (10th ed. 2014). We tend to think of software as non-autonomous and self-driving cars are more than just software. 127 See Artificial Intelligence, supra note 123. 128 AI News: Artificial Intelligence Trends and Leading Stocks, INV. BUS. DAILY, https://ww
yk4x [https://perma.cc/XP3R-F69G]; Chris Giarratana, How AI is Driving the Future of Au-tonomous Cars, READWRITE (Dec. 20, 2016), https://readwrite.com/2016/12/20/ai-driving-fu
ture-autonomous-cars-tl4/ [https://perma.cc/B5SE-X9PL]. 132 See Giarratana, supra note 131. 133 NEV. REV. STAT. § 482A.020 (repealed 2013).
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
1032 NEVADA LAW JOURNAL [Vol. 19:3
system is truly autonomous, from one end being the air-conditioning system to
the other end being the system navigation, we would call the whole thing “au-
tonomous.” Thus, the generality in which “AI” is defined is excusable because
the goal is to create a composite system within the context and scope of the
statute.
C. Louisiana
A Louisiana statute focuses on defining “electronic agents” in regards to
autonomation, but notably, an annotation for this statute contemplates the fu-
ture developments of AI and how there could be a relevant distinction between
weak AI and strong AI.134 Under Louisiana’s property, contracts, and financial
services code, Louisiana law accounts for electronic commerce and the elec-
tronic agents that could be used to carry out such transactions.135 The Louisiana
Uniform Electronic Transaction Act includes an annotation demarcating be-
tween different types of electronic agents:
5. “Electronic agent.”
(a) This definition establishes that an electronic agent is a machine. As the term
is used in this Chapter, it is limited to the function of a tool. The effect on the
party using the agent is addressed in the operative provisions of this Chapter
(e.g., Section 2614).
(b) An electronic agent, such as a computer program or other automated means
employed by a person, is a tool of that person. As a general rule, the employer of
a tool is responsible for the results obtained by the use of that tool since the tool
has no independent volition of its own. However, an electronic agent, by defini-
tion, is capable within the parameters of its programming, of initiating, respond-
ing or interacting with other parties or their electronic agents once it has been
activated by a party, without further attention of that party.
(c) While this Chapter presupposes that an electronic agent is capable of per-
forming only within the technical strictures of its preset programming, it is con-
ceivable that in the future, electronic agents may be created with the ability to
act autonomously, and not just automatically. That is, through developments in
artificial intelligence, a computer may be able to “learn through experience,
modify the instructions in their own programs, and even devise new instruc-
tions.” If these developments occur, the courts may construe the definition of
electronic agent accordingly, to recognize such new capabilities.136
This statute’s annotation articulates the differences between weak and
strong AI. The former is explained in sections (a) and (b) as nothing more than
a tool, and such, that tool is confined to its functions outlined by its program-
ming parameters.137 The annotation goes as far as to assign liability to the user
of the computer, even if the computer communicates with other parties or
agents autonomously.138 Here, the statute recognizes the difference between au-
tomation and autonomy even though in this circumstance they go hand in hand,
yet the user is still liable because the program is acting within its own parame-
ters. However, the annotations go one step further and acknowledge strong AI
as a possibility.
Section (c) recognizes that AI can learn and develop as it has experiences,
exhibiting functions of machine learning.139 Furthermore, the annotation cate-
gorizes AI as something that can devise new instructions and change its own
programing, further highlighting the similarities between human and AI.140 The
Louisiana statute also explicitly states that it may need to modify this statute as
technology advances, demonstrating how this technology might evolve.141 This
statute denotes legislative intent for how to qualify computer agents as either
strong AI or weak AI machines. This definition is at least on the right track,
although it might still run into the same problem present within the Nevada
statute as only being applicable to a specific context.
Although these examples are not all of the definitions out there, they pro-
vide a good lay of the land. Definitions can either be overly broad and include a
wide array of systems, or they can be field specific and define AI for a specific
application. The difference in approach can lead to different problems. Creating
a definition for each application is burdensome and will nevertheless skirt the
issue of defining AI. With the ever-expanding applications of AI, each field or
subset of laws will need to include a definition for AI for that particular use.
However, if jurisdictions wish to define AI specifically for each application,
then every new application still begs the question, “does this specific instance
qualify as ‘AI’?” While this is not necessarily problematic, it is certainly cum-
bersome and frustrating. A general definition for AI can be applied across the
board to all fields if written properly.
IV. METHODS FOR WRITING A GENERAL DEFINITION
Creating a functioning legal definition is difficult enough, but there is the
added issue of who should determine definitions and regulations—the courts,
legislature, or other government agencies?142 As Justice Alito mentioned, the
legislature might have done the research and formulated a particular law or def-
inition, so they might be in a good position for their given circumstances.143 As
mentioned above, Nevada’s legislature only sought fit, at least at the moment,
to define AI in terms of its function in self-driving cars.144 Statutes can be
138 Id. 139 Id.; see also Vincent, supra note 2. 140 See discussion supra Parts I–II. 141 LA. STAT. ANN. § 9:2602 (2018). 142 See Scherer, supra note 5, at 356. 143 See Brown v. Entm’t Merchs. Ass’n, 564 U.S. 786, 806 (2011) (Alito, J., concurring). 144 See supra Section III.B.
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
1034 NEVADA LAW JOURNAL [Vol. 19:3
preemptory or reactionary, but nonetheless, policy makers need to understand
AI, specifically strong AI. Without comprehension there cannot be a sufficient
legal definition.145 Furthermore, this requires understanding how “AI” func-
tions as a phrase to be able to define it pragmatically.
Despite the common use of words or phrases, “shared words need not have
shared meanings,” and “legal systems can ascribe whatever meaning(s) they
desire to the words that they borrow, inherit, or invent.”146 Although the words
“artificial intelligence” are shared words, they do not have a shared meaning.
Part of the legislature’s and judiciary’s role is to define the “meanings of words
to reflect the unique priorities, preferences, and goals of a judicial, political, or
social system.”147 These goals might include legal regulation or economic con-
siderations that will certainly arise.148 For certain words, it is possible to discern
a common meaning from its variant meanings;149 however, because “AI’s”
meaning is all over the place, a common meaning is difficult to discern, and
perhaps unnecessary.150 “Traffic,” for instance, is an ordinary term that has a
couple definitions, but 18 U.S.C § 2320 defines it purposely for the section.151
This is standard—codes and acts are going to have these definitions out of ne-
cessity.152 Thus, stipulating a definition can rid “AI” of the ambiguity in which
the phrase is used.
Ridding terms of ambiguity is essential because terms have functions and
effects.153 The major function is to tell people what a term or phrase means in
particular contexts, or put simply, to clarify.154 However, the second function is
to empower, or to give a particular status to something or someone.155 These
145 See supra Parts I–II. 146 Thomas O. Main, The Word Commons and Foreign Laws, 46 CORNELL INT’L L.J. 219, 222, 225 (2013) (“Because words have more than one meaning, we can discern from those variant meanings what we might call a common meaning.”). 147 Id. at 222. 148 See id. at 219; see also discussion supra Part I. 149 See Main, supra note 146, at 222. 150 See Price, supra note 11, at 1005–06 (“But other definitions of familiar but context-dependent terms impose unexpected restrictions on the scope of word meaning.”). 151 18 U.S.C. § 2320(f)(5) (2012) (“[T]he term ‘traffic’ means to transport, transfer, or oth-erwise dispose of, to another, for purposes of commercial advantage or private financial gain, or to make, import, export, obtain control of, or possess, with intent to so transport, transfer, or otherwise dispose of”). In this example, “traffic” is used in a sense far different than “traffic violation.” 49 U.S.C. § 31301 (2012). This is similar to how AI was defined for autonomous cars. See Section III.B. 152 See Section III.B. 153 See Price, supra note 11, at 1026–51. 154 Id. at 1026 (“Statutory definitions even more emphatically illustrate the two levels on which legal language operates. On the one hand, statutory definitions tell us what terms mean in particular contexts; they are meant to clarify the message conveyed by the legisla-ture in the statute’s normative clauses.”). 155 Id. at 1027 (“On the other hand, legislative definitions can also empower; they may con-fer a particular status on individuals, entities, or situations, and invest those agents and states of affairs with obligations, benefits, privileges, and rights.”).
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
Spring 2019] ARTIFICIAL INTELLIGENCE 1035
two functions are critical for understanding how “AI” should be statutorily de-
fined. The particular issue when keeping such functions in mind is that “our
language is built by its speakers,” and so meanings change overtime as people
use terms.156 Compare how AI was being used to play chess as opposed to
where AI is now—no longer following an algorithm but making its own deci-
sions.157 Another reason for clarity and precision in statutory definitions is due
process.158 A law must be clear enough to be easily understandable by every
citizen to ensure it is just,159 and with the conflation of different types of sys-
tems, it could be difficult to identify one’s legal status, responsibility, or expo-
sure. Hence, it is all about proving context to eradicate ambiguity,160 and there
is plenty of ambiguity in technology, let alone with intelligent systems.
Altogether, defining terms without a valid and well-identified intention
risks complication.161 “[D]efinition[s] can provide information and direction on
how a particular term is to be understood.”162 No matter the purpose, defini-
tions ought to be simple and succinct.163 The defining process is never com-
plete, and unforeseen situations and changes are inevitable. Thus, it is especial-
ly important to devise cogent definitions for terms which describe subjects that
are continuously evolving.164 Even if there is no static definition for AI, “it
ought to simplify clarify the legislative intention.”165 The determination must
be made whether it is a prescriptive or descriptive definition, in order to deter-
mine the definition’s scope. Thus, it is critical for the judiciary and legislative
branches to formulate the definition specific.
156 Id. at 1031 (“Underlying nearly all discussions of language and communication is the idea of community. Our language is built by its speakers.”); see also Peter R.A. Gray, The Language of Statutes: Laws and Their Interpretation: Lawrence M. Solan (2010), 19 INT’L J. SPEECH, LANGUAGE & L. 135, 139 (2012) (book review). 157 See Section II.A–II.C. 158 See Price, supra note 11, at 1031 (“Due process considerations certainly require that generally applicable statutes should have a meaning that is clear ‘enough’ to those to whom it applies.”). 159 Id. 160 Id. (“Context’s role in clarifying meaning is obvious: context allows disambiguation . . . .”). 161 Id. at 1051 (“Defining without a valid and well-identified purpose risks complicating the statute and creating the potential for unforeseen questions unrelated to the statute’s pur-pose.”). 162 Id. at 1052. 163 Id. (“No matter what purposes it serves, however, the definition ought to be simple and succinct in its statement.”). 164 Id. at 1053 (“[T]he legislative drafter would do well to realize that definition is a process that is always incomplete.”). 165 Id.
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
1036 NEVADA LAW JOURNAL [Vol. 19:3
A. Ambiguity & Descriptors
Statutes commonly define ordinary terms that could be ambiguous.166 In
“Wagging, Not Barking: Statutory Definitions,” by Jeanine Frazier Price, Pro-
fessor Price outlines two modes of defining terms: ambiguous terms and de-
scriptor terms,167 both of which are applicable to artificial intelligence. Ambig-
uous terms, such as “artificial intelligence,” have a scope issue, and the point of
a statutory definition is to pick out or stipulate which of the many definitions
apply.168 However, when a term functions as descriptor, it refers to a group or
set that shares well-recognized characteristics (for example, “child” does not
pick out one thing, but rather a group that resembles one another).169 This dis-
tinction between ambiguous types and descriptor types is important. Ambigu-
ous terms have one referent. Conversely, descriptor terms have more than one
referent sharing common characteristics, rather than the term itself.170
Applying this distinction to “artificial intelligence” is critical because it ei-
ther has one referent, what artificial intelligence actually is, or the shared char-
acteristics for everything called “artificial intelligence.” Though the latter still
requires those characteristics to be defined, it can explain why many different
things are commonly referred to as “artificial intelligence.” However, there are
ways to attempt to define these types of terms.171 One option for a definition is
to list what is included in the scope.172 For example, “livestock” is defined as
“cattle, swine, and lambs.”173 This mode of defining is helpful in outlining what
specifically falls under a given definition. Here, the definition has the option of
listing off systems, software, programs, or modes of computer functions that
qualify as “artificial intelligence.” This would be a descriptive definition, as
opposed to a prescriptive definition that focuses on key characteristics or fea-
tures that are present in all artificial systems. Both methods can be suitable for
defining artificial intelligence so long as the drafters have an understanding of
the characteristics of artificial intelligence, and the distinction between strong
AI and weak AI.
166 Id. at 1005 (“Sometimes those definitions clarify what sense of the word is intended, and place limits on the term’s application for purposes of the statute.”). 167 Id. at 1006–07. 168 Id. at 1005–07; see also discussion supra Parts I–II. 169 See Price, supra note 11, at 1007. 170 See id. 171 See id. at 1010–11. 172 See id. at 1010. 173 See 7 U.S.C. § 1635a (5) (2012).
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
Spring 2019] ARTIFICIAL INTELLIGENCE 1037
B. Descriptive Definition
Descriptive definitions describe a term in a way that typically reflects “the
actual use of [the] word or phrase.”174 The list could include radically different
things, but they are unified by common characteristics. For example, “nuclear
science” is defined as including “(1) nuclear science; (2) nuclear engineering;
(3) nuclear chemistry; (4) radio chemistry; and (5) health physics.”175 It is pret-
ty obvious by this definition that health physics is different than nuclear engi-
neering. Price argues that descriptive definitions create “fuzzy categories” in
which they are not bound by necessary or sufficient conditions, but rather by
the closeness of resemblance to the original instance.176 So for AI, this categor-
ical approach allows for its definitions to have loose conditions based on com-
mon characteristics.177 Thus, a descriptive definition for artificial intelligence
would be a list of several uses of artificial intelligence, such as machine learn-
ing, natural language, virtual agents, and reactive machines.178
For example, “artificial intelligence includes (1) Reactive machines, (2)
Limited Memory machines, (3) Theory of Mind systems, and (4) Self-
awareness systems, or include other systems that utilize autonomous deep
learning.” This definition would rely on the current scientific consensus of the
existing types of AI systems, and the current existing types may not consider
certain attributes. This is not inherently problematic, but it creates equality be-
tween different systems that have contrasting functions and different levels of
autonomy.
Furthermore, this method has a few problems. One major problem is that
this method shifts to the debate to which systems count as AI systems. Second,
the “fuzzy category” allows for argumentation, which may not necessarily be a
bad thing, but it is less definitive. Although this might complicate the legisla-
ture’s work, it does allow for there to be a clear distinction between strong AI
and weak AI, and thus, this method has some merit.
174 Price, supra note 11, at 1010–11 (“Descriptive definitions, not surprisingly, describe the meaning of a term. And, that description usually reflects or depends upon the actual use of a word or phrase, although the definition itself may enlarge or contract ordinary usage.”). 175 42 U.S.C. § 16532 (2012). 176 Price, supra note 11, at 1011 (“Descriptive definitions often depict either a cluster of at-tributes of what is defined or typical uses of the term.”). 177 Id. (“That definition could encompass radically different objects and instances . . . but each of those instances share some characteristics . . .”). 178 Arend Hintze, Understanding the Four Types of Artificial Intelligence, GOV’T TECH. (Nov. 14, 2016), http://www.govtech.com/computing/Understanding-the-Four-Types-of-Arti
ficial-Intelligence.html [https://perma.cc/WT6R-LC64]; see also RUSSELL & NORVIG, supra note 56, at 2–3.
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
1038 NEVADA LAW JOURNAL [Vol. 19:3
C. Prescriptive Definition
Comparatively, prescriptive definitions can outline a set of necessary con-
ditions that must be satisfied to fall within the scope of the definition.179 For
example, the definition of “aircraft” outlines necessary conditions that must all
be satisfied in order to be an aircraft: it must be “a civil, military, or public con-
trivance invented, used, or designed to navigate, fly, or travel in the air.”180
This definition outlines several necessary conditions that must be satisfied to
qualify as an aircraft, albeit with multiple ways to satisfy each condition. Pre-
scriptive definitions are precise but rely on the designation of essential charac-
teristics, which can establish features that distinguish members of the set.181
Artificial intelligence would then be defined by its characteristics (such as
those mentioned above) and could specifically outline two separate definitions
for strong AI and weak AI.182 Using this method, a definition for strong AI
might look like: “artificial intelligence is a system, program, software, or algo-
rithm that acts autonomously to think rationally, think humanely, act rationally,
act humanely, make decisions, or provide outputs.”183 To change this definition
to a weak AI definition, the words “acts autonomously” can be substituted for
something similar such as “follows instructions.” This method allows for a def-
inition with some flexibility to include various or key characteristics; however,
the list relies on determining the appropriate set of characteristics.
These two methods offer different benefits for the term or phrase in ques-
tion.184 For terms with multiple meanings, descriptive definitions can specify
which sense applies, but this can change the meaning by what is included or
excluded.185 The definition is going to be under-or over-inclusive,186 and this
creates an additional consideration when determining whether prescriptive or
descriptive methods should be used.187 Clearly these are not the only two meth-
ods of creating statutory definitions, but they are two methods that are condu-
cive to defining complex terms. Either of these methods accomplish the goal of
subjecting the definition of “AI” to normative rules,188 and having a precise
179 See Price, supra note 11, at 1012 (“Prescriptive definitions may consist of a set of condi-tions, compliance with each one of which is necessary to fall within the definition’s scope . . . .”). 180 18 U.S.C § 31(a)(1) (2012). 181 See Price, supra note 11, at 1013 (“They may assign meaning to a term by designating a broader class and then establishing features that distinguish members of the defined term.”). 182 See id. at 1009–13. 183 See supra Section II.A–II.B. 184 See Price, supra note 11, at 1013–15. 185 Id. at 1015 (“For terms of many meanings, descriptive definitions can quickly and effi-ciently indicate which sense of the word or phrase applies.”). 186 See Scherer, supra note 5, at 373. 187 See Price, supra note 11, at 1010–17. 188 Id. at 1018 (“Falling within a statutory definition results in becoming subject to norma-tive rules.”).
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
Spring 2019] ARTIFICIAL INTELLIGENCE 1039
definition can narrow uncertainty as well as promote predictability.189 Either of
these methods can accomplish the goal of removing ambiguity and making a
statutory distinction between strong AI and weak AI. Thus, the definition can
account for the unique characteristics of a new technology.
CONCLUSION
Despite the lack of consensus as to what artificial intelligence is, it is being
developed at an alarming rate and will without a doubt be one of the world’s
most significant developments.190 There is going to be a legal vacuum within
the next decade or two where courts and legislatures will struggle to get a han-
dle on how to deal with AI.191 Rather than being reactive, courts, legislatures,
or any legal body for that matter, will be better off navigating these new waters
with some guidance from a statutory definition.
However, these new waters are going to be best traversed if we understand
the distinction between autonomous systems and systems following a given al-
gorithm or specified procedure. Current definitions only seem to regulate and
portray an understanding of weak AI. Typically, because technological ad-
vancements are unforeseeable to a certain extent, and because most regulations
are put in place after problems arise, definitions clarifying these advancements
will come after. But with AI, the advancement—at least to the development of
autonomy—is foreseeable. This autonomy will accelerate the development of
AI and autonomy at a staggering rate.192
We will greatly benefit from systems that are able to operate outside of our
input and free from our human biases. Artificial intelligence will undoubtedly
move in the direction of autonomy exclusive of humans because computers are
just better at making decisions than people.193 Of course arguments can be
made about whether or not these decisions should be left to machines; however,
with proper regulation and statutory guidelines, we can at least understand how
the current technology is different than past technology. Nevertheless, people
will happily allow an AI system to diagnose them if it means surviving, or hap-
pily embrace an investment system able to make them money. Companies are
189 See REED DICKERSON, THE FUNDAMENTALS OF LEGAL DRAFTING 145 (2d ed. 1986). 190 See supra Part I. 191 See Scherer, supra note 5, at 391–92. 192 Louis Columbus, 10 Charts That Will Change Your Perspective on Artificial Intelli-gence’s Growth, FORBES (Jan. 12, 2018, 1:55 AM), https://www.forbes.com/sites/louiscolum
h/#ed8f65947583 [https://perma.cc/TKH5-6QDU]; The Evolution of Artificial Intelligence: AI’s Coming of Age, UBS, https://www.ubs.com/microsites/artificial-intelligence/en/ai-comi
ng-age.html [https://perma.cc/9R9A-YNGF] (last visited Mar. 16, 2019); see also supra Parts I–II. 193 Lance Whitney, Are Computers Already Smarter Than Humans?, TIME (Sept. 29, 2017), http://time.com/4960778/computers-smarter-than-humans/ [https://perma.cc/7FJU-6UAB]; see also supra Part I.
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
1040 NEVADA LAW JOURNAL [Vol. 19:3
already investing billions of dollars in this industry, and rather than being reac-
tive to the situations and consequences that arise, it is far better to be prepared
for the inevitable legal, legislative, and judiciary discussions to come. One way
to do this is by having a legal definition for “AI” to untangle the current ubiqui-
tous use of the phrase now.
The definition needs to make clear the distinction between strong AI and
weak AI, as different legal questions arise out of the distinction.194 These defi-
nitions can be done for particular applications, such as the Nevada statute for
driverless cars.195 However, a general definition that could be applied across
multiple applications could be beneficial so long as the distinction between
strong AI and weak AI is understood, even if they end up being treated similar-
ly. Nevertheless, the definitions need to be flexible because of expansive and
innumerable ways in which artificial intelligence can be utilized.196 The mean-
ings of words are determined by use,197 and one word can have multiple mean-
ings; further, meanings can change over time.198 “Artificial intelligence” al-
ready has a lot of uses, and it is necessary for the legal field to stipulate a
workable definition in order to save itself time and resources. “New ideas and
concepts spawn new words,”199 and AI is a field that will undergo rapid and
exponential advancement, even beyond what we can consider now. This is why
there needs to be language in place to handle this influx. However, ambiguity
and vagueness are inescapable; so, a definition must be revisited whenever nec-
essary, and specifically with this type of advancement, revising the language
should happen often.200 Because cases can turn on the definition of a term,
courts need to settle on a definition and give it legal weight and authority.201 I
have offered some examples of definitions, and I do not pretend to think they
are anywhere near perfect; however, they demonstrate how one could write def-
initions that can account for autonomous machines.
Once again, I am not saying that the law needs to treat strong AI and weak
AI differently, just that the distinction needs to be understood because they are
in fact very distinct from one another. To go back to the parent-child analogy,
we legally treat minors differently than adults.202 It did not have to be this way,
there could be a world where minors are just as legally accountable as adults,
but we understand how minors, and if taken to the extreme, how babies, are not
194 See discussion supra Section II.D. 195 See supra Section III.B. 196 See discussion supra Parts I, IV. 197 Main, supra note 146, at 228 (“These terms are defined so that we can use them in later parts to explore the overlap and interaction of meanings that result when different legal sys-tems use the same words.”). 198 Id. at 230–31. 199 Id. at 231. 200 See id. at 234. 201 Id. at 233–34 (“Social conventions recognize and accept judicial authority to declare the meaning of words, albeit for a limited purpose and for a particular discourse community.”). 202 See Cunningham, supra note 116, at 277–78.
19 NEV. L.J. 1015, MARTINEZ 5/28/2019 10:48 AM
Spring 2019] ARTIFICIAL INTELLIGENCE 1041
autonomous, and thus we should treat them differently. We could only do this
if we understand the difference in intelligence between someone who is adult
and someone who is a minor. The United States draws its line in the sand at 18
years of age, and I draw the line in the sand at autonomy.203
An upcoming machine age is on its way and it will be led by AI.204 This
technology will change the world, and that means the legal profession should
be ready and prepared to account for those changes. There are two outcomes
from AI, a utopian one and a dystopian one.205 Either the plot of Terminator
unfolds, or we learn to live with machines and respect their artificial capaci-
ties.206 Depending on our understanding and capabilities to deal with AI, legal
or otherwise, we can hope to achieve the better of the two outcomes.
203 I understand that my line is not a bright red one, but I think that could have its benefits and allow for some malleability, just as sometimes the United States charges children as adults. See Nicole Scialabba, Should Juveniles Be Changed as Adults in the Criminal Justice System?, A.B.A. (Oct. 3, 2016), https://www.americanbar.org/groups/litigation/committees/c